Introduction

This work is highly inspired by https://paintschainer.preferred.tech/ and http://www.cse.cuhk.edu.hk/~ttwong/papers/colorize/colorize.pdf.
The goal is to show how easy it is to train your own coloring network by using Pytorch, OpenCV, and the public Danbooru dataset.

How to use it

1. Clone the repo, then run "python irodoru.py" in command line. Note that you might need to install CUDA and all the missing dependencies, if so, please skip to the next section.
2. Use any browser to open "http://localhost:8080" and start playing!

Installing the necessary dependencies

1. It's strongly recommended to use a conda environment for running this repo (e.g. https://www.anaconda.com/download/)
2. Download and install CUDA 10 from https://developer.nvidia.com/cuda-downloads
3. To install PyTorch:
   • Windows: conda install pytorch torchvision cuda100 -c pytorch
   • Linux: conda install pytorch torchvision cuda100 -c pytorch
4. For OpenCV: conda install -c conda-forge opencv
5. Use command "pip install flask flask-socketio pillow numpy scipy scikit-learn eventlet" if you miss some of the libraries.

Preparing the dataset

The entire training is done over the danbooru 2017 512px dataset, which can be acquired using the torrent files:
https://www.gwern.net/docs/anime/danbooru2017-sfw512px-torrent.tar.xz
Once you have the images, make sure to remove all the 8-bit grayscale ones, perhaps with the supplied "python train.py --clean "

Training the models

The workflow is as follows:

1. Train the draft model with "python train.py --draft [danbooru directory] [model directory]"
2. Generate the dataset for refinement model with "python train.py --generate [danbooru directory] [draft image directory] --model [draft model]"
3. Train the refinement model with "python train.py --refine [draft image directory] [model directory]"

Future work